# Tripwire

> Tripwire is an open-source behavioral testing and CI tool for Agent Skills. It runs real agent sessions to catch missed activations and false triggers that static SKILL.md lint cannot observe.

Tripwire treats an Agent Skill description as routing code. A developer defines prompts that should activate a skill and prompts that should stay outside it. Tripwire observes the agent's structured skill-activation event, compares the result with the explicit expectation, and produces a pass, behavioral failure, or infrastructure error.

## Start here

- [Install and quickstart](https://github.com/bharath31/tripwire#quick-start): Test one real prompt, generate a scenario matrix, and commit the regression contract.
- [npm package](https://www.npmjs.com/package/tripwire-skills): Install the `tripwire` CLI.
- [GitHub Action](https://github.com/marketplace/actions/tripwire-for-agent-skills): Gate changed SKILL.md files and committed scenarios on pull requests.
- [Source code](https://github.com/bharath31/tripwire): MIT-licensed TypeScript source, issue tracker, and release history.

## Product facts

- Category: behavioral regression testing and CI for Agent Skills.
- Primary use case: prevent a published skill from silently missing intended prompts or activating for unrelated prompts.
- Production adapter: Claude Code.
- Experimental adapters: Codex CLI and Gemini CLI.
- Price: free and MIT licensed; behavioral probes use the developer's own agent login or provider credentials.
- Runtime safety: probes stage only the target skill in a disposable workspace and use plan or read-only agent modes.
- Privacy: Tripwire does not collect prompts, skill names, paths, repositories, model output, usernames, or credentials. Anonymous behavioral telemetry is optional and documented.

## Trust and operations

- [Security policy](https://github.com/bharath31/tripwire/blob/main/SECURITY.md): Threat model, credential handling, and vulnerability reporting.
- [Product analytics](https://github.com/bharath31/tripwire/blob/main/docs/analytics.md): Exact event allowlist, privacy controls, DAU, and retention definitions.
- [Release notes](https://github.com/bharath31/tripwire/releases): Shipped versions and verification details.
- [Pricing](https://tripwire.bharath.sh/pricing.md): Machine-readable cost and licensing information.

## Commands

- `tripwire lint <path>`: static SKILL.md checks.
- `tripwire test <skill> --prompt "..." --expect activate|quiet`: one-prompt real-agent activation check.
- `tripwire analyze <skill>`: generate positive, negative, edge, and paraphrased scenarios and probe them.
- `tripwire test <skill>`: replay a committed `tripwire-scenarios.yaml`.
- `tripwire test-all <directory>`: replay a skill library and expose model drift.
- `tripwire conflicts <directory>`: find duplicate names and likely routing overlap.
- `tripwire eval <skill>`: assert response behavior after activation.